On Thu, Jan 31, 2013 at 10:15:44PM -0500, Chris Burroughs wrote: > On 01/31/2013 08:55 AM, Willy Tarreau wrote: > > This one has everything needed, transfer-encoding: chunked specifies the > > size so the connection can stay alive. > > > >> But responses from haproxy still closed with either http-server-close or > >> http-pretend-keepalive set still close the connection. I admit these > >> results are confusing and I feel like I'm missing something. > > > > You definitely need to have both options set, and no "httpclose" at all. > > > > Got it working with (A) no httpclose, (B) http-server-close, (C) > http-pretend-keepalive, (D) enabling keep-alive on the backend. > > What I was originally looking for was a way to avoid (D). I'm not sure > if that's supposed to be possible, but everything looks stable so > success as far as I'm concerned.
This is a design bug in tomcat's HTTP implementation, so we have no other choice here : the server considers that keep-alive==HTTP/1.1 and close==HTTP/1.0, and when it's in close mode it doesn't send content-length nor does it chunk so the message is only delimited by an end of connection. The "http-pretend-keepalive" workaround was developped precisely because of such servers. In your case it's not a problem since haproxy will close the connection anyway. Still it's annoying to have to do that. I have some vague memories of someone here reporting this on the tomcat ML, resulting in a fix one or two years ago, but I may confuse with something else. Maybe you should experiment with newer versions ? What you need is just the server to emit content-length with every response, including close ones, and to fall back to chunked encoding when content-length is unknown. Regards, Willy
